DeepSeek Founder's Investor Call Reveals AGI Roadmap and Compute Plans

A four-hour transcript of DeepSeek founder Liang Wenfeng's investor meeting has circulated widely, covering the company's AGI roadmap, open-source strategy, product trade-offs, hardware investments, and organizational culture. The transcript reveals DeepSeek's clear technological roadmap and unique philosophy on survival and business.

Confirmed

Core Strategy and Product Trade-offs: Liang stated that DeepSeek is solely focused on AGI, with consumer and enterprise products being mere stepping stones and byproducts. He believes Coding Agents will materialize before general agents or vertical AIs. The next-gen model's primary task is to improve the efficiency of "self-training" and emphasize continual learning. He estimates AI's long-term value could reach 10% of global GDP, but a single entity should not attempt to monopolize it.

Hardware and Domestic Compute: Liang revealed that compute spending in 2026 will likely not exceed 20 billion RMB. As of early June, the company had about 20,000 H-equivalent compute units. He explicitly stated a desire to buy chips at reasonable prices rather than investing in self-developed chips. The V4 official release has been delayed (originally targeted for June), and training a 150B active parameter model is scheduled for late 2026 to April 2027. Huawei's 950 supernode is considered a viable replacement for GB200/GB300, with domestic AI chip adaptation facing only production capacity bottlenecks.

Management Philosophy and Business Paradigm: Authors analyzing the transcript highlighted Liang's principle of "restraint." @AlchainHust noted that the real takeaway isn't a mysterious AGI blueprint, but simple principles like knowing what not to do, avoiding overexpansion, and maintaining team stability. Regarding open source, Liang views it as a business strategy: if model pricing allows cost recovery in about 10 months, third-party deployments become unprofitable, ensuring open source doesn't harm revenue.

Why it matters

This transcript not only answers external questions about DeepSeek's commercialization and compute infrastructure but also showcases a rare business paradigm: using the deterministic cash flow from quantitative trading to support the uncertainty of AGI exploration. This highly restrained approach, which drives model scaling through algorithmic cost reduction, provides a differentiated survival model for the domestic LLM landscape.

2026-07-21 ~ 2026-07-23 · 37 related posts

• 100-trillion-parameter models may soon orchestrate multi-agent systems, as inference demand keeps growing — sudoraohacker · 2026-07-21
• DeepSeek’s Liang Wenfeng says AGI, open source and cost discipline are the company’s core strategy — elsewhere别处发生 · 2026-07-22
• [source] DeepSeek’s Liang Wenfeng Says Coding Agents Come Before General Agents and Vertical AI — oran_ge · 2026-07-23
• Liang Wenfeng says product is just a step on the way to AGI — oran_ge · 2026-07-23
• Liang Wenfeng’s cash-flow-backed AGI bet is a business model few founders can copy — yongqianme · 2026-07-23
• Leaked DeepSeek recording surfaces Liang Wenfeng’s views on AGI and management — vista8 · 2026-07-23
• [source] Liang Wenfeng’s four-hour DeepSeek talk is really a manifesto of restraint — AlchainHust · 2026-07-23
• Zhipu chief says open source is the company’s core agenda, and restraint is key to survival — teortaxesTex · 2026-07-23
• DeepSeek’s Wenfeng calls 10-month hardware payback a “reasonable profit” — teortaxesTex · 2026-07-23
• ByteDance-linked discussion says AI ARR could hit hundreds of millions, maybe $1B — teortaxesTex · 2026-07-23
• DeepSeek says open source won’t hurt revenue if models recoup costs in 10 months — teortaxesTex · 2026-07-23
• DeepSeek says its open-source models are the same ones it deploys internally — teortaxesTex · 2026-07-23
• A long-term AGI roadmap puts agents first, then continual learning, then embodied AI — teortaxesTex · 2026-07-23
• OpenAI-style AGI-first strategy says consumer and enterprise are just byproducts — teortaxesTex · 2026-07-23
• Liang Wenfeng says AI could capture 10% of GDP, but no one should monopolize it — max_paperclips · 2026-07-23
• Liang Wenfeng says DeepSeek is chasing AGI, not commercialization, in a 4-hour investor meeting — 创业邦 · 2026-07-23
• Moonshot figure’s surprise at Daya’s exit exposes a harder line on talent retention — teortaxesTex · 2026-07-23
• Moonshot says world models are outside its AGI mainline and not worth priority — teortaxesTex · 2026-07-23
• DeepSeek says it has 20,000 H-equivalent units and wants to spend into GPUs fast — teortaxesTex · 2026-07-23
• [source] DeepSeek’s Wenfeng says 2026 compute spend may stay under RMB 20B — teortaxesTex · 2026-07-23
• China AI speaker says 2026 compute spend may stay under RMB 20B — teortaxesTex · 2026-07-23
• Chinese companies are likely to become AI’s biggest producer, quote says — teortaxesTex · 2026-07-23
• Domestic AI chip adoption has no ecosystem blocker, only a production bottleneck — teortaxesTex · 2026-07-23
• Liang Wenfeng backs open source as a commercial strategy for AI — sebkrier · 2026-07-23
• Huawei 950 supernode is described as a GB200/GB300 substitute, but capacity is still tight — teortaxesTex · 2026-07-23
• A quoted interview says the company still hopes to buy chips, not make them — teortaxesTex · 2026-07-23
• DeepSeek’s next model should add continual learning before chasing bigger gains — teortaxesTex · 2026-07-23
• DeepSeek Aims for Continual Learning in Next-Gen Models to Drive Down Costs — teortaxesTex · 2026-07-23
• DeepSeek says cheaper training lets it build larger models on the same compute budget — teortaxesTex · 2026-07-23
• DeepSeek says formal work should stay under half the week, with the rest left for research — teortaxesTex · 2026-07-23
• DeepSeek says affordability, not max returns, is the point of its model strategy — teortaxesTex · 2026-07-23
• DeepSeek founder says restraint is the strategy on the road to AGI — vaibhavbetter · 2026-07-23
• Liang Wenfeng says DeepSeek builds models for internal usefulness first — teortaxesTex · 2026-07-23
• Liang Wenfeng says 16,000 Huawei 950 GPUs are only enough for current-gen training — teortaxesTex · 2026-07-23
• Model roadmap hints at a late-June release and a possible 150B run — teortaxesTex · 2026-07-23
• V4 GA is delayed, and 150B-active training is now penciled in for late 2026 to April 2027 — teortaxesTex · 2026-07-23

1 near-duplicate retellings: teortaxesTex